Detached One-Off

A detached one-off is a component instance you deliberately broke away from its master definition so you could tweak it for one specific use case. In tools like Figma this usually means right clicking and selecting detach instance. What was once a living reference to your tokens and variants becomes a static group of layers with hardcoded values. The colors no longer pull from your semantic token set. The spacing stops respecting your scale. The states that should live as named variants now exist as random layer styles. The machine reading your file cannot tell which version is authoritative. It sees the clean component on your components page and the detached rebel living in the dashboard mockup as equally valid. Both inform what the AI generates next. That is how one lazy edit becomes the new normal at scale.

It is not a proper new variant added to the main component with clear properties like size small or state error. It is not an override that keeps the instance linked to its source definition. It is not an intentional one time mock that gets deleted after the review ends. Those approaches preserve the integrity of the system and feed clean data to any tool. A detached one-off does the opposite. It creates a permanent fork that lives inside your production files and teaches bad habits to anyone or anything that reads them. It is the design systems version of technical debt that compounds with interest every time an AI tool points at the file.

Consider what happened at Ramp in Q1 2025. Their design system included a strict set of surface tokens for cards with four elevation levels. A product designer detached a payment summary card two days before launch to accommodate a new progress indicator. The change involved a 2 pixel adjustment to the border radius and a hardcoded shadow value that did not exist in the token set. The file went to engineering as is. Weeks later when the team adopted Figma AI for rapid iteration on new billing flows the generated designs showed four different card styles mixed together. One followed the tokens. One copied the detached version. Two invented hybrids. The resulting developer handoff contained 23 separate bugs traced directly to the conflicting definitions. The design systems manager later calculated that the single detached one off cost the team nine full engineering days to correct.

A similar disaster played out at Notion during their 2024 Q3 AI features push. The team had excellent component variants for their database blocks including named states for hover focus and selected. One designer detached several instances to create a new empty state for the AI generated views. Those detached blocks carried different typography scales and padding that contradicted the official tokens. When product managers started using Claude Design to spin up additional templates the outputs contained inconsistent block styling throughout. Notion eventually had to dedicate a full time designer for a month to hunt down and eliminate every detached instance across 47 different team files. Their post mortem listed detached one offs as the number one cause of design drift in AI accelerated workflows.

The same pattern bit GitHub while updating Primer in 2025. A marketing designer detached a modal component to add extra vertical spacing for a campaign page. That 32 pixel override never made it back into the main component. When the team pointed their internal AI documentation generator at the file it began outputting modals in two competing sizes across every new feature doc. The systems team now runs a nightly script that flags detached instances in red and blocks any file containing more than two from being used as reference data.

Reach for a detached one off only in the earliest exploration phase of a brand new pattern your current system cannot express. Set a hard 48 hour timer to integrate the learnings back into the main component as a named variant with proper token references. That is the only acceptable use. Do not reach for them when handing off to developers. Do not reach for them in any file destined for AI consumption in 2026. Do not reach for them because adding a new variant property feels tedious. That extra three minutes of work prevents weeks of downstream fixes once the machine starts amplifying the contradiction. Do not reach for them in long lived project files or design system documentation. The longer they survive the more legitimate they appear to both human collaborators and machine interpreters. In 2026 the cost of leaving one behind is no longer measured in single screens. It is measured in the hundreds of generated outputs that embed the contradiction permanently.

Design governance at scale demands zero tolerance. Teams at Shopify and Atlassian run automated checks that surface every detached instance during pull requests for design files. The checks require explicit tickets explaining why the detachment was necessary and a plan to restore the link or convert it to a variant. This level of discipline turns machine readability from an aspiration into a guarantee. Without it your tokens and variants become suggestions instead of rules. The machine does not guess intent. It reads what exists and ships it at full speed.

Every detached one off erodes the structure that AI tools depend on. Tokens lose their authority. Component definitions fragment. States become ambiguous. The machine does not shrug and pick the best one. It amplifies whatever it sees most often or most recently in the file. That is why cleanup must happen before any generation pass. Reattach what you can. Delete what you cannot. Rebuild from the official sources using named variants and token references. The five minutes you spend now saves hours of inconsistency later when Figma AI Claude Design or the next tool treats your entire library as training data.

Detached one offs turn your best design decisions into background noise while amplifying every lazy shortcut you ever took.